> The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
The New York Times lawsuit isn't purely about training on content - it's about memorization. They included examples of OpenAI's models being able to reproduce NYT articles entirely from the weights.
There's also stuff about RAG in there - about OpenAI (and Microsoft) features that run web searches and then summarize the results. I've not been following developments closely, I'd be interested to know how that part of the lawsuit is shaking out so far.
Right, it's not just that they saw the thing but also how they use the knowledge later.
To demonstrate this "what you do next matters" part, assume I get a book from the library and I read it, and I have perfect memory. So far, no problem, even with my fantastical memory. However what I do next matters:
A. I quote several short pieces of the book in a written critique. Fair use.
B. I use what I know to create a statistical breakdown of the author's favorite words and phrases. That's fine too.
C. I retype my own book which is 90% identical to the original and I publish it for money. Nuh-uh. Legal problems abound.
We can debate whether or not these companies are crossing the line, but the line does depend on what they do with the information, not just how they got it.
The reproduction argument was always the weaker of the arguments. The whole point of news is that it's new, if you ask ChatGPT for the current front page story that's not gonna be in the weights. They were able to contrive a single example of a very well known piece, but the number of people who will use that to circumvent the paywall is zero.
The Bartz v. Anthropic settlement wasn't about fair use, it was about pirating ebooks. The judge in that case said that the training on scanned, purchased books was fair use - Anthropic started doing that in 2024.
And it supports my summary here. Anthropic did "win" on the fair use issue, but lost on the pirating.
Note that I said "The Bartz v. Anthropic settlement wasn't about fair use" - and that's accurate, the settlement aka the $1.5bn was about the pirated books, even though the initial lawsuit was about fair use.
> Code spitters are cooked, let's be honest here and accept the fate so everyone can move on.
I'm guessing "code spitters" here means people who consider their job to be manually writing code based on requirements produced by others.
Personally, I think the role of "developer" has already expanded way beyond that. If you're a technology expert who understands what is and is not possible to build, how it can be built, how to prompt that thing into existence, verify that the implementation is good, then get that implementation safely into production, you're not "cooked" at all.
For once I actually agree with Simon, and fortunately the percentage of engineers interested in software architecture and business concerns is so low already that I doubt they’ll have trouble in the job market. Pure “coders” were not safe from layoffs and PIPs even before LLMs, in my experience, since their jobs could so easily be outsourced and offshored.
It describes a team spending 30x payroll on tokens - $3,000,000 in token spend per month - building a multiplayer game where incidents and bug reports from users are passed directly to Fable 5/5.1 agents who then act on them and deploy the changes to production.
"We consider it an outage if we don't ship every hour"
It's worth reading because it's either a complete disaster or a glimpse of the future, or quite possibly both.
well, i did go check out the company spawn from the article. its certainly interesting how hands off their process is and how much they ship, but holy shit the games were fucking horrible. there was a direct WoW ripoff (right down to the assets) that was unplayable, a racing simulator where my controls stopped working 10 second in, and some sort of RPG that i started and just never loaded. so i will take what they are saying with the heaviest possible grain of salt.
Yes exactly this is my standard for everything now. You can spend an unbelievable amount of tokens doing all kinds of things that are interesting as processes in their own right.
at the end of it: is anything of value delivered to customers? the company might say "well users make the games with AI, we just serve them and host the multiplayer." okay, but again, what value was created? they brag about being one of anthropic's top customers; congrats you are handing VC money to anthropic. why should i care? theres no answer for why anyone should care, because to them the system and clout is the product, the product doesnt actually matter.
at least thats my read on it. i dont know how else to understand whats going on here.
Oh it's Jacob - I used to work w/ him on his acquired startup, Pluto. He was very AI-pilled; I'm not surprised at the results. I remember wading through the AI-generated Python he had written. I think Pluto had something like 500k or 700k lines after 18mo. Everything was duplicated; there was no reuse of anything. Dozens of features, which mostly worked, not all did.
As a product that I would regularly use, Pluto missed the mark for me. As a SWE in 2023, I couldn't meet Jacob's mentality. Today, it's how a lot of us ship software.
I didn't love the product but I have to respect the process: he was able to sell his company. For some, what he did is success.
That actually sounds really cool. Assuming it has access to a well-tuned predictive analytics mechanism that allows it to derive that changing the game in this way would increase satisfaction and revenue from you and other players like you, sufficiently to offset development costs and dissatisfaction to others, then it really should.
We are already - as in today - doing this with friends while playing Valheim, the process is semi-automatic with me in the middle as the admin, though.
We make 10+ gameplay changes a day in our custom kitchensink mod that steers everything. 70%+ are bugfixes (we have plenty of mods that conflict/are buggy), the the rest are QoL improvements, small new features and rebalancing. If we see some playstyle is too strong or boring, we spice it up the same or the next day. Mobs too weak/strong? Update the same day. A ton of performance fixes with actual in-game benchmarking. All this while I keep playing.
I built sort of an automatic modding pipeline (including public release and all) with parallel test game clients + servers used by LLMs to develop and test the changes and then merge before releasing a new mod version. We have had very few issues, I'm still shocked how reliable this is. They follow guidance well and test everything in real game clients on wipeable clones of our live world and our actual characters. Josh has a bugged chest at his place? The LLM will reproduce it, develop a patch, yes, but then also log in as him to his actual in-game base and test the fix using game inputs + OCR. At the end I get a .html page report on each change with screenshots and technical details if I ask for them. I don't care if it takes 4h to do a simple fix because it's not my time and game stability is critical. And up to 8 of them can work in parallel, so there isn't a bottleneck caused by excessive testing/re-testing.
I think this will be the future of gaming for many sandbox/easily moddable games. It's a bit too technical at the moment for regular players to handle the updates and models' quirks, but we're only going up from here. Just crazy.
I could definitely see potential in an async co-op game based on this idea. A sim harness, some pbt to avoid broken states, and something like a modified nomic ruleset might get pretty far.
whoa like an ai calvinball. Or, you compete to get the right to one prompt that changes the rules/physics/map etc. Whoever gets to the genie's lamp first …
This is amazing and somewhat insane. It's more like they are building an organism with an immune system than a code base. Interesting that they say only Fable works for them.
I’m building the same but with a malleable SaaS. Basically when you sign up, you get the software deployed and an agent that can control that deployment and code.
But the process is slow. The user wants to add a feature to their app so they ask the AI. The AI has access to a terminal where it can clone, fix, and push. This takes time and isn’t instant.
Once AI can abstract this part, software will change forever. It’s what this post alludes to, does code, PR, etc need to reinvented for this?
Once so, no one would ever use cookie cutter software because you could build your own and maintain it using just an agent.
Yes but now that cookie cutter software can be the starting point. I’m not saying to build it from scratch. There’s probably an open source alternatives to the software you are trying to replace. Use that directly or tell AI to clone it and build it in a stack you understand.
Before AI, the businesses that used Zapier and automation were able to pull ahead of businesses that were still doing things manually. A business that has a custom built software with an agent that can change it on the go will provide even a greater advantage.
This is a complete and utter disaster. They are making spam games. Like miniclips but for spam. A semi functional website to make what is basically youtube playables.
And it's devouring tens of millions. I know what tens of millions can make in the video game industry, it can keep a small team of good devs and designers and artists making something special that's _really_ enjoyed by tens of thousands of people.
We don't want more games, we want more great games. This will never, ever, make great games.
Look on ye mighty, and despair. I really don't care how this sausage was made because the only thing you need to know is that it shouldn't ever have been made.
As a (light) gamer for two decades who has been reading about the gaming industry for the past few months, I cannot understate how hard it is to create a good game, let alone a great game. So much money was spent/wasted on hopeless endeavors.
Was thinking of that exact article while I was reading this, we must be following the same folks. Highly recommend folks reading it, even if it feels insane in some regards it's a fascinating look at what the limits of this stuff could be and I found myself with a lot of new ideas to consider after reading it. Interesting times.
Edit - after writing "we must be following the same folks" I realized that you are, in fact, Simon Willison - I saw this article because you posted about it on X, haha.
This works as we've been doing it since Opus 4.5 (Nov 2025) with every user request, feature and bug. It's the future without a doubt. The feedback loops evolve and get stronger depending on how well you fine-tune or train your system to optimize on its own.
That’s the dumbest thing I’ve ever read. Spaffing millions per month on this slop is the kind of thing people suffering from AI psychosis do. There is no way you need “thousands of agents” to build any product.
The more agents you throw at something, the more they collide. Those collisions generate more slop, which requires further slop processing agents.
> A default that made Savi think longer on every message, landed at 10:22pm and reverted by 10:40pm: that's not right, undo it it's too slow and not jazz.
I giggled on that one. It feels like children playing having a job. If they pay that much for tokens they have to have at least 80k customers paying 50$ a month. Seems huge for something like that. Or they have a lot of investors. Anyway does not feel like something going anywhere.
This from that tweet stood out to me: "In general, we've seen that OpenAI's models game their own checks"
In the past when working on agentic stuff I've observed models gaming checks aggressively, to the point that when we were validating log files to see what the agent had done, the agent responded by editing the log files to pass validation. I can only imagine how bad that kind of behavior is at scale, and it sounds like they ran into it.
It this point, a whole lot of teams have been leaning into AI-first, not-entirely-reviewed code since around March this year.
That means we are already 6-8 months in, and not far from having a full year of experience of what that actually looks like, albeit weirdly influenced by the increase in model capabilities during that period.
I'm looking forward to those retrospectives. Did this turn out to work OK for the teams doing it? What were the biggest failings and/or surprises? With hindsight, did this produce good quality software?
I think the biggest shift is in predictability. Will something be done today, next week, or next month - nobody knows. Will we converge on a bug free state? not sure.
So far AI works great at fast moving work which you will probably recode or throw away in a month or two anyway (e.g. a startup) but is weak at maintaining a large code base that you rely on. At least part of this is cultural, as the cost of misalignments between team members can be quickly translated into broken software.
Software robustness, at least in terms of "has the app crashed on any client machines today", is achievable. It's just expensive enough to achieve that most people (rightly or not) don't prioritize it, and the more moving parts (like third party libraries or language runtimes) your application has the harder it is to achieve. Maybe AI will lower the cost of achieving that.
When I worked on Guild Wars every server or user crash went to a shared inbox and was bucketed so we could identify and act on new crashes. IIRC we maintained very, very low production crash rates and could prove that many of the crashes were due to hardware faults.
Similarly, Mozilla tries to analyze and act on every crash report they get in the field. If a webpage visited by even a small % of the customer base was causing crashes, the team worked to try and fix the crash to deliver stability. If it was a high-traffic webpage it got treated as a high-priority task - I remember Farmville breaking was basically a 3-alarm fire back when it was extremely popular.
"No bugs at all" is an even higher bar, but it probably is possible for simpler applications with the use of stuff like theorem provers. I'd love to see someone try it in production but I think the return on investment for that is probably super super negative.
I operate on a "no known bugs" basis with quite a bit of effort invested to try to find new, unknown bugs.
The idea that every crash isn't acted on is kind of a crazy one, to me. Not every crash is actually actionable in e.g. a mobile app, but every one should at least get an eyeball in my opinion.
This, to me anyway, is the bare minimum of "caring about software quality". If you do this regularly you'll have very solid code, and it's interesting that I have the perception at least that Guild Wars was a quality product.
There's another layer, where you do that for subthreshold issues like UI inconvenience, etc, but that's bonus, I think.
For all the talk about AI making teams super productive, I don't see much talk about AI improving software quality.
It's a shame, though, because I believe there is a lot of potential there.
I mean, quality is easier with the right tooling (monitoring, linters, property-based tests, theorem prover...). But, given limited resources, it can feel difficult to prioritize investing in quality when business is pressing devs to ship, ship, ship.
But now the cost of writing custom tooling has changed. These custom tools might be sloppy because vibe-coded but we wouldn't have them otherwise! Or we might finally have decided to make that investment after too much time suffering from their absence. Now the cost/benefit changed, so why not invest now?
> But now the cost of writing custom tooling has changed. These custom tools might be sloppy because vibe-coded but we wouldn't have them otherwise
There’s a lot of good tools out there that are not used, either because of ignorance of their existence or lack of expertise. An hour of config can pay real dividend later.
And sometimes you just use some unix utilities and hack a bash script while testing things out.
After 15 years on the field, I still lack a lot of expertise. Many teams do, as you say.
LLM can save hours of doc reading for non-critical tools that I would otherwise not use. Its also pretty good at "hacking a bash script" in order to test things out. Not everyone is confortable writing bash scripts.
Have been at it since Nov 23rd 2025 with a first commit with Opus 4.5 - it was rough at first but after about year well worth it. You can also tell where so many just get stuck in the mud due to lack of direct experience with all the engineering silos (from bare metal sys admin to QA eng silos outside of standard software engineering core) it takes to roll out a new kind of technology company while scaling a new kind of commercialization model, also managed by AI and AI native from the ground up.
I expect Proton's existing employees may be a particularly bad audience for a "let the AI worry about security" message from the top, given that they are more likely to be a) security experts and b) people who care enough about security to actively seek out a job at Proton.
> “I’ve given him unlimited tokens because he’s going to refactor code,” Patel told Fortune at a small roundtable dinner. “So I’m like, you can go, you can go nuts.”
I think that Fortune story is better than the Tom's Hardware one. It ties the Cisco anecdote into a larger trend of some people deliberately hiring younger, less experienced developers for their AI projects - because they are more likely to be "AI native".
Yes, I’m sure the Indian man at Cisco acquired a translation company founded by another Indian man because he’s young and AI native, not for any other reason.
Sadly, I think the label "AI naïve" is better applied to the many experienced professionals who refuse to engage with AI systems as a matter of principle - or who tried out ChatGPT in 2024, saw it make howlingly obvious errors, and wrote off the entire category.
I frequently find myself in conversations with people who are resistant to AI and turn out to have a very skewed idea of what it can and cannot do.
> or who tried out ChatGPT in 2024, saw it make howlingly obvious errors, and wrote off the entire category.
Or they were reading your comments and articles and couldn't reproduce any of the fantastic results that you were claiming, even back then. Things have indeed changed since the first Opus release about a year ago and LLMs are good at programming now. But still, experienced professionals are not so keen to sign off 1000s of lines of code that they never read, take responsibility for it etc. Also, it is a different job and it will take some time to adjust. Can you blame them?
Moreover, other trades have their own guilds or are otherwise protected through regulation. Can you blame programmers for fighting to keep doing things as they were doing it until recently?
The personal energy usage impact of LLMs remains tiny, even for projects that use millions of tokens. Skip one 20 minute drive and you should be covered for weeks, if not months.
I've recently talked to people at Anthropic who are very confident that they have effectively solved the prompt injection risks that cause lethal trifecta attacks with their latest models.
And we have to be honest: both agents will misuse users and the other way around. So you are asking Anthropic for a backing that will also get stress test by meatspace. I do like the idea: a prepaid or restricted amount credit card for agentic purposes, with the novelty being that amounts above threshold have some kind of protection within consumer law.
LLMs regularly confuse my messages with their own. They'll say "you said xyz" when it was actually them. I heard the same principle can be used to bypass security restrictions, so maybe it's still an issue. (Or do they only get confused about authorship when security isn't at stake?)
I have literally never seen this and it doesn't make much since because the turns are flagged with metadata. It should be impossible for Claude to confuse your text for its own. I know because I use this metadata all the time. I've built hooks and processes against it.
Contexts get compacted, texts get copy-pasted, tasks get transferred between agents, humans assume ownership of LLM generated content without reading it, etc. etc.
Yes all user error and avoidable using good processes but I have not seen any organization care about this so far. The carelessness of AI adoption is crazy. People are not aware of the risks.
Isn't prompt injection just one facet of the trifecta? Like, I just don't know how it's remotely possible to have a non-deterministic block box ever deterministically know whether your credit card has been compromised or not. Even if you were hacked, how would you even know?
Like the guy who put his SSN on the side of a truck and asked people to steal his identity?
I like it!
IIRC he had his identity stolen a bunch.
----
> In 2005, Davis and Robert Maynard Jr. cofounded LifeLock,[10] a personal fraud protection company.[11] Davis was CEO of LifeLock since its founding. As part of a marketing campaign to promote his company's credit monitoring services, Davis posted his own social security number[12] on billboards and commercials. As a result, he was a victim of at least 13 cases of identity theft between 2007 and 2008, belying LifeLock's claims that its services protect consumers against identity theft and fraud.[13] The Federal Trade Commission imposed a $12 million penalty on the company for deceptive advertising, likening Davis and LifeLock marketers to "con artists."[14]
Or the user being an idiot. Saying this somewhat sarcastically, but there’s always someone who says “Go buy me a Rolex and have it shipped to my house” and then didn’t realize that could be $25,000
> The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
The New York Times lawsuit isn't purely about training on content - it's about memorization. They included examples of OpenAI's models being able to reproduce NYT articles entirely from the weights.
There's also stuff about RAG in there - about OpenAI (and Microsoft) features that run web searches and then summarize the results. I've not been following developments closely, I'd be interested to know how that part of the lawsuit is shaking out so far.
reply